debian-mirror-gitlab/doc/development/database_review.md

Database Review Guidelines

This page is specific to database reviews. Please refer to our code review guide for broader advice and best practices for code review in general.

General process

A database review is required for:

• Changes that touch the database schema or perform data migrations, including files in:
  • db/
  • lib/gitlab/background_migration/
• Changes to the database tooling. For example:
  • migration or ActiveRecord helpers in lib/gitlab/database/
  • load balancing
• Changes that produce SQL queries that are beyond the obvious. It is generally up to the author of a merge request to decide whether or not complex queries are being introduced and if they require a database review.

A database reviewer is expected to look out for obviously complex queries in the change and review those closer. If the author does not point out specific queries for review and there are no obviously complex queries, it is enough to concentrate on reviewing the migration only.

It is preferable to review queries in SQL form and generally accepted to ask the author to translate any ActiveRecord queries in SQL form for review.

Roles and process

A Merge Request author's role is to:

• Decide whether a database review is needed.
• If database review is needed, add the ~database label.
• Prepare the merge request for a database review.

A database reviewer's role is to:

• Perform a first-pass review on the MR and suggest improvements to the author.
• Once satisfied, relabel the MR with ~"database::reviewed", approve it, and reassign MR to the database maintainer suggested by Reviewer Roulette.

A database maintainer's role is to:

• Perform the final database review on the MR.
• Discuss further improvements or other relevant changes with the database reviewer and the MR author.
• Finally approve the MR and relabel the MR with ~"database::approved"
• Merge the MR if no other approvals are pending or pass it on to other maintainers as required (frontend, backend, docs).

Distributing review workload

Review workload is distributed using reviewer roulette (example). The MR author should then co-assign the suggested database reviewer. When they give their sign-off, they will hand over to the suggested database maintainer.

If reviewer roulette didn't suggest a database reviewer & maintainer, make sure you have applied the ~database label and rerun the danger-review CI job, or pick someone from the @gl-database team.

How to prepare the merge request for a database review

In order to make reviewing easier and therefore faster, please take the following preparations into account.

Preparation when adding migrations

• Ensure db/structure.sql is updated as documented.
• Make migrations reversible by using the change method or include a down method when using up.
  • Include either a rollback procedure or describe how to rollback changes.
• Add the output of both migrating and rolling back for all migrations into the MR description.
  • Ensure the down method reverts the changes in db/structure.sql.
  • Update the migration output whenever you modify the migrations during the review process.
• Add tests for the migration in spec/migrations if necessary. See Testing Rails migrations at GitLab for more details.
• When high-traffic tables are involved in the migration, use the with_lock_retries helper method. Review the relevant examples in our documentation for use cases and solutions.
• Ensure RuboCop checks are not disabled unless there's a valid reason to.
• When adding an index to a large table, test its execution using CREATE INDEX CONCURRENTLY in the #database-lab Slack channel and add the execution time to the MR description:
  • Execution time largely varies between #database-lab and GitLab.com, but an elevated execution time from #database-lab can give a hint that the execution on GitLab.com will also be considerably high.
  • If the execution from #database-lab is longer than 1h, the index should be moved to a post-migration. Keep in mind that in this case you may need to split the migration and the application changes in separate releases to ensure the index will be in place when the code that needs it will be deployed.

Preparation when adding or modifying queries

• Write the raw SQL in the MR description. Preferably formatted nicely with pgFormatter or paste.depesz.com.
• Include the output of EXPLAIN (ANALYZE, BUFFERS) of the relevant queries in the description. If the output is too long, wrap it in <details> blocks, paste it in a GitLab Snippet, or provide the link to the plan at: explain.depesz.com.
• When providing query plans, make sure it hits enough data:
  • You can use a GitLab production replica to test your queries on a large scale, through the #database-lab Slack channel or through chatops.
  • Usually, the gitlab-org namespace (namespace_id = 9970) and the gitlab-org/gitlab-foss (project_id = 13083) or the gitlab-org/gitlab (project_id = 278964) projects provide enough data to serve as a good example.
• For query changes, it is best to provide the SQL query along with a plan before and after the change. This helps to spot differences quickly.
• Include data that shows the performance improvement, preferably in the form of a benchmark.

Preparation when adding foreign keys to existing tables

• Include a migration to remove orphaned rows in the source table before adding the foreign key.
• Remove any instances of dependent: ... that may no longer be necessary.

Preparation when adding tables

• Order columns based on the Ordering Table Columns guidelines.
• Add foreign keys to any columns pointing to data in other tables, including an index.
• Add indexes for fields that are used in statements such as WHERE, ORDER BY, GROUP BY, and JOINs.

Preparation when removing columns, tables, indexes, or other structures

• Follow the guidelines on dropping columns.
• Generally it's best practice (but not a hard rule) to remove indexes and foreign keys in a post-deployment migration.
  • Exceptions include removing indexes and foreign keys for small tables.
• If you're adding a composite index, another index might become redundant, so remove that in the same migration. For example adding index(column_A, column_B, column_C) makes the indexes index(column_A, column_B) and index(column_A) redundant.

How to review for database

• Check migrations
  • Review relational modeling and design choices
  • Review migrations follow database migration style guide, for example
    • Check ordering of columns
    • Check indexes are present for foreign keys
  • Ensure that migrations execute in a transaction or only contain concurrent index/foreign key helpers (with transactions disabled)
  • If an index to a large table is added and its execution time was elevated (more than 1h) on #database-lab:
    • Ensure it was added in a post-migration.
    • Maintainer: After the merge request is merged, notify Release Managers about it on #f_upcoming_release Slack channel.
  • Check consistency with db/structure.sql and that migrations are reversible
  • Check queries timing (If any): Queries executed in a migration need to fit comfortably within 15s - preferably much less than that - on GitLab.com.
  • For column removals, make sure the column has been ignored in a previous release
• Check background migrations:
  • Establish a time estimate for execution on GitLab.com. For historical purposes, it's highly recommended to include this estimation on the merge request description.
  • If a single update is below than 1s the query can be placed directly in a regular migration (inside db/migrate).
  • Background migrations are normally used, but not limited to:
    • Migrating data in larger tables.
    • Making numerous SQL queries per record in a dataset.
  • Review queries (for example, make sure batch sizes are fine)
  • Because execution time can be longer than for a regular migration, it's suggested to treat background migrations as post migrations: place them in db/post_migrate instead of db/migrate. Keep in mind that post migrations are executed post-deployment in production.
• Check timing guidelines for migrations
• Check migrations are reversible and implement a #down method
• Check data migrations:
  • Establish a time estimate for execution on GitLab.com.
  • Depending on timing, data migrations can be placed on regular, post-deploy, or background migrations.
  • Data migrations should be reversible too or come with a description of how to reverse, when possible. This applies to all types of migrations (regular, post-deploy, background).
• Query performance
  • Check for any obviously complex queries and queries the author specifically points out for review (if any)
  • If not present yet, ask the author to provide SQL queries and query plans (for example, by using chatops or direct database access)
  • For given queries, review parameters regarding data distribution
  • Check query plans and suggest improvements to queries (changing the query, schema or adding indexes and similar)
  • General guideline is for queries to come in below 100ms execution time
  • If queries rely on prior migrations that are not present yet on production (eg indexes, columns), you can use a one-off instance from the restore pipeline in order to establish a proper testing environment. If you don't have access to this project, reach out to #database on Slack to get advice on how to proceed.
  • Avoid N+1 problems and minimalize the query count.

Timing guidelines for migrations

In general, migrations for a single deploy shouldn't take longer than 1 hour for GitLab.com. The following guidelines are not hard rules, they were estimated to keep migration timing to a minimum.

NOTE: Note: Keep in mind that all runtimes should be measured against GitLab.com.

Migration Type	Execution Time Recommended	Notes
Regular migrations on db/migrate	3 minutes	A valid exception are index creation as this can take a long time.
Post migrations on db/post_migrate	10 minutes
Background migrations	---	Since these are suitable for larger tables, it's not possible to set a precise timing guideline, however, any single query must stay below 1 second execution time with cold caches.